Evaluate ScyllaDB as a potential DB to improve write-heavy PushPull

[hackerwins]

What would you like to be added:

Yorkie uses a PushPull-based synchronization mechanism to exchange document changes between clients and the server. During each PushPull operation:

• Push: The client pushes local changes to the server.
• Pull: The server responds with changes the client has not yet seen.
• During the Pull phase, Yorkie also updates checkpoints and versionvectors to MongoDB, recording which changes the client has received.

As a result, compared to typical web services, Yorkie exhibits a significantly higher write load, and overall performance is tightly coupled to the database's write throughput and durability guarantees.

Current performance

In our current setup (Apple M1 Max / 32GB) using standalone Yorkie + MongoDB 8.0, we achieve:

• 3,000 concurrent connections with presence updates in a single document (skewed workload)
• 10,000 concurrent connections with an even distribution across documents
• https://github.com/yorkie-team/yorkie/tree/main/test/k6#run-the-load-test

Related work: Discord and ScyllaDB

• Discord originally built their message system on MongoDB but transitioned to ScyllaDB to handle scale and improve write performance.
• In a benchmark report by Tractian, ScyllaDB demonstrated 10x better throughput in write-heavy workloads compared to MongoDB.

Given that Yorkie is similarly write-intensive due to its PushPull operations, evaluating ScyllaDB as a backend database might offer significant performance improvements.

Proposal

• Implement a ScyllaDB backend interface in the Yorkie database layer
• Compare performance (especially write latency and throughput) against MongoDB 8.0
• Use the same presence load test to measure improvements
• Decide whether it is viable for high-scale workloads

References

• MongoDB 8.0 performance improvements: https://medium.com/@jsupskills/mongodb-8-0-ready-to-use-with-reduce-memory-and-faster-speed-85a16adc8686
• Tractian’s benchmark and migration to ScyllaDB: https://www.scylladb.com/2023/04/03/mongodb-vs-postgres-vs-scylladb-tractians-benchmarking-and-migration/

Why is this needed: